A Multi-View Deep Learning Framework for EEG Seizure Detection

A Multi-View Deep Learning Framework for EEG Seizure Detection

Abstract: The recent advances in pervasive sensing technologies have enabled us to monitor and analyze the multi-channel electroencephalogram (EEG) signals of epilepsy patients to prevent serious outcomes caused by epileptic seizures. To avoid manual visual inspection from long-term EEG readings, automatic EEG seizure detection has garnered increasing attention among researchers. In this paper, we present a unified multi-view deep learning framework to capture brain abnormalities associated with seizures based on multi-channel scalp EEG signals. The proposed approach is an end-to-end model that is able to jointly learn multi-view features from both unsupervised multichannel EEG reconstruction and supervised seizure detection via Spectrogram representation. We construct a new auto en coder based multi-view learning model by incorporating both inter and intra correlations of EEG channels to unleash the power of multi-channel information. By adding a channel-wise competition mechanism in the training phase, we propose a channelaware seizure detection module to guide our multi-view structure to focus on important and relevant EEG channels. To validate the effectiveness of the proposed framework, extensive experiments against nine baselines, including both traditional handcrafted feature extraction and conventional deep learning methods,

are carried out on a benchmark scalp EEG dataset. Experimental results show that the proposed model is able to achieve higher average accuracy and f1-score at 94.37% and 85.34%, respectively, using 5-fold subject-independent cross validation, demonstrating a powerful and effective method in the task of EEG seizure detection.

Existing system: Recently, numerous studies have been conducted on automatic EEG seizure detection. In general, the task of automatic EEG seizure detection can be divided into two main stages: feature extraction and classification. In particular, feature extraction is key, since its aim is to characterize distinctive EEG patterns, which directly affect the performance of the classifier. Towards this end, most of the existing work employs expert handcrafted features to detect EEG seizures, which can be broadly classified into two categories, namely temporal domain features and spectral-domain features. The design of these handcrafted features, though different in various aspects, often involves a significant amount of expert knowledge to derive interpretable representations from data. For instance, Fourier analysis is a typical signal processing tool to extract frequency contents from signals. Since brain abnormality is often reflected by increased amplitude and Frequency changes in EEG signals, the results of different Fourier transform modifications could be used as features for EEG seizure detection. In addition, principal component analysis (PCA) , a statistical procedure to explore distinctive features using orthogonal transformation, has also been frequently employed for reducing the redundancy in feature space to improve the accuracy in EEG seizure detection .

Proposed system: Various methods for handcrafted feature extraction for EEG seizure detection have been proposed. Several researchers propose to combine expert handcrafted features with a classifier, e.g., support vector machine (SVM) or neural networks (NN), to build a multi-stage EEG seizure detection system. Firstly, different types of features have been extracted in the time domain, such as phase , entropy, and other statistical measurements . Secondly, various spectral features have also been widely used. Both the Fourier transform and wavelet transform can be exploited

for this purpose. Thirdly, there has been a growing interest in introducing timefrequency image descriptors to EEG seizure detection. These methods attempt to capture the non stationary behavior of EEG signals by integrating both image and signal related features.

Advantages: This study focuses on the problem of offline seizure detection. Below we briefly discuss some representative EEG seizure detection methods in this area.

Firstly, different types of features have been extracted in the time domain, such as phase, entropy, and other statistical measurements. Secondly, various spectral features have also been widely used.

Numerous techniques exploiting this idea have been applied for multi-channel EEG detection, such as PCA individual component analysis (ICA) , Laplacian eigen maps, and multi-dimensional scaling . By selecting a smaller subset of useful features, the classifier performance could be improved.

Disadvantages: Since the evaluation task belongs to a classification problem, we use F1-score and Accuracy to evaluate our model. Moreover, the receiver operator characteristic (ROC) curve and the precision-recall (PR) curve are plotted to illustrate the diagnostic ability of the seizure detector.

Epilepsy is characterized by unprovoked seizures associated with sudden irregular neuronal discharges in the brain and can result in other health problems, even death. Electroencephalogram (EEG) is one of the commonly used methods to monitor, detect, and diagnose epileptic seizures due to its affordability and high temporal resolution.

Modules: Epileptic seizure: EPILEPSY is the fourth most common neurological disorder, and there are approximately 50 million people affected by epilepsy worldwide. Epilepsy is characterized by unprovoked seizures associated with sudden irregular neuronal

discharges in the brain and can result in other health problems, even death.

Electroencephalogram (EEG) is one of the commonly used methods to monitor, detect, and diagnose epileptic seizures due to its affordability and high temporal resolution. EEG signals often contain multiple channels which measure brain electrical activities through several electrodes placed in different locations of the brain. According to previous work, EEG signals are usually collected with electrodes placed on the scalp or exposed surface of the brain, referred to as scalp EEG and intracranial EEG, respectively. Most studies of EEG seizure detection prefer to utilize the scalp EEG since it can be acquired with non-invasive technologies. In clinical practice, EEG readings require highly-trained professionals in neurology. However, the long term EEG visual inspection is extremely time-consuming and laborious for physicians to identify characteristic patterns of seizures from multiple channels. Thus, automatic EEG seizure detection has garnered great interest among researchers in neuro-informatics.

Handcrafted Feature Extraction: Various methods for handcrafted feature extraction for EEG seizure detection have been proposed. Several researchers propose to combine expert handcrafted features with a classifier, e.g., support vector machine (SVM) or neural networks (NN), to build a multi-stage EEG seizure detection system. Firstly, different types of features have been extracted in the time domain, such as phase, entropy, and other statistical measurements. Secondly various spectral features have also been widely used. Both the Fourier transform and wavelet transform can be exploited for this purpose. Thirdly, there has been a growing interest in introducing time-frequency image descriptors to EEG seizure detection. These methods attempt to capture the non stationary behavior of EEG signals by integrating both image and signal related features. Considering the fact that a typical EEG system often contains multiple electrode channels, the dimension of features may grow substantially with the increase in the number of channels.

Multi-view Learning: With the recent advancement in artificial intelligence, multi view learning has drawn significant attention in a wide range of applications. The goal of multi-view learning is to learn features from multiple perspectives to improve the

generalization performance. Traditionally, canonical correlation analysis (CCA) and its kernel extensions have been commonly adopted in early studies of multiview learning technologies. Several advanced algorithms have been incorporated to improve the learning performance, including topic learning, sparse coding, nonnegative matrix factorization, and Markov networks. Recently, inspired by the success of deep neural networks, some hybrid models that combine multi-view learning with deep learning have been proposed to learn shared representations.

Channel-aware Module Analysis : In this subsection, we present the performance improvements of our models when using the proposed channel aware module. Specifically, we compare with the results using the conventional “full-channel� training strategy, referred to as cSSDA (Full), cConvA (Full), mConvA (Full), and mSSDA (Full), respectively.

The comparison results are reported in Fig. 5. From the figure, we can see that the performance of all the models increases on average two percentage points in terms of Accuracy and F1-score when adopting the channel-aware module. It means that by adding channel wise sparsity and selectivity constraints, the models can be trained to focus on important and relevant channels, and thus enhance the feature representation of multi-channel data input. As shown in Table II, the proposed channel-aware module supports the effectiveness of different multi-view deep learning structures, which achieve good performance in EEG seizure detection.